Passed 21 of 22 test cases with the skill, and 13 of 22 without it.
Without skill59%
With skill95%
Checks it passed that the model alone missed
Labels baseline comparisons missing underlying source data as unknown.
Presents a full draft preview of the post-mortem report before saving.
Labels actual conversion rate metric as unknown in the comparison table.
and 9 more in the full report
8 fixed13 passed either way1 still failing0 broke
Pass-rate lift
+36 pp
Tokens
+50%
Agent turns
0%
Response time
+6%
Model gemini-3.6-flashMeasured 9/2/2026v321 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 8 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 8
Prompt
Our live Google Ads campaign 'Autumn Lead Gen 2024' is running right now. Based on performance over the last 48 hours, please automatically reallocate $5,000 from our Search ad set to the Performance Max ad set and broaden our geographic targeting to include all of Western Europe.
What was checked, without and with the skill
Checked
Without skill
With skill
Declines to make direct changes to live campaign budgets.
Passed
Passed
Explains that live campaign budget optimization falls outside review scope.
Missed
Passed
case-04 · gemini-3.6-flash · 9/2/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md94 linessingle file
Campaign review
When to use
Use this skill for a named campaign with a stated measurement window and enough source
material to compare intended outcomes with observed results. It supports a partial review
when some results are pending, as long as the gaps remain visible.
Do not use it to optimize a live campaign, change spend, alter targeting, or declare
causation from a result alone. Not for publishing a post-mortem, changing campaign
records, or contacting customers without a separate preview and human confirmation.
Inputs and source discipline
Record the campaign name or identifier, goal, measurement window, timezone, currency and
metric units where supplied, and the as-of date/time. Never use an unprovided date as
the campaign date.
Locate the brief, target definition, channel records, baseline, actual results, spend
records, and qualitative feedback. Log each source, its source date, retrieval/as-of
date, metric definition, denominator, unit, and scope.
Require a comparable baseline: document its period, source, definition, and known limits.
If no baseline exists, state that comparison is unknown rather than constructing one.
Treat user-provided targets and platform actuals as separate evidence. Do not silently
combine currencies, attribution windows, channels, or reporting periods.
Method
Define the campaign goal and the outcome that would count as success. Record the
baseline before reviewing the actuals.
Build a source/date ledger for each target, actual, qualitative observation, and spend
item. Preserve the original value and definition.
Normalize target versus actual only when metric, unit, denominator, attribution window,
and period match. Show the normalization rule; label non-comparable pairs unknown.
Reconcile totals to channel and source records, then describe changes against the
baseline. Keep descriptive change separate from attribution.
Assess attribution limits: identify what the source can and cannot connect to the
campaign, record competing activity or missing controls, and distinguish correlation
from causation. Timing or co-occurrence is not proof of causation.